We developed an automated methodology for real-time validation of hydrometric data in a sewer network. Our methodology uses real-time validated data to optimise system management and non-real-time data to evaluate day-to-day performance.
Two approaches can be used to validate and correct hydrometric data; the choice depends on the number of level gauges present in a system. In single gauge systems, univariate filtering is used to smooth data. For example, frequency filtering systematically eliminates values corresponding to frequencies higher than a predetermined threshold frequency. In systems with several gauging stations-duplex, triplex, or multiplex systems-the multivariate filtering method proposed here can be used to validate data series from each gauge. Material redundancy in duplex or higher order systems makes it possible to detect a deficient gauge, using a decision rule to set aside erroneous readings before averaging accepted values. Part of the underlying principle of this methodology is heavier reliance on gauges that give readings consistent with previous and subsequent validated values in a given series. Thus isolated positive or negative variations within a series are eliminated if corresponding variation values at other gauges are more consistent. To evaluate persistence, a reading is compared to a value predicted by an autoregressive (AR) model calibrated by the previous validated reading.
This filtering technique constitutes an intelligent alternative to the frequency filtering method mentioned above. In more practical terms, it compares the deviation of an AR model prediction from a measured value with the deviation of the same AR model prediction from a value estimated by a regressive model at other stations in the network. Among the values measured and estimated by the regressive model, the one nearest the AR model prediction is retained.
Our methodology also relies on analytical redundancy generated by direct measurement of flow and hydrological simulation. More precisely, the deviation of the AR model prediction from the measured value is compared with the deviation of the same AR model prediction from a value obtained from a hydrological simulation model. Among measured and simulated values, the one nearest the AR model prediction is retained. To allow consideration of nonstationary models and to avoid the well-known bias of the least squares method, the Kalman filter is used to identify the parameters of the AR model.
The methodology we propose employs three models. The first generates analytical redundancy using hydrological modelling. An autoregressive model is then used to predict future runoff rate values. Finally, a voting process model is used to compare measured and simulated values.
The proposed methodology was tested on the Verdun sewer system in Quebec with successful results. Two types of artificial disturbance of the measured hydrograph were created: white noise was added to measured values and disturbances of large amplitude and various forms were introduced. The methodology produced the initial values and performance criteria were conclusive. Thus on-site testing confirms that this approach allows completely automated detection and correction of most anomalies. Flood peaks were neither underestimated nor overestimated, and total runoff volumes were retained.
Validation, redundancy, flow, measurement, Kalman filter, autoregressive, real time, sewer.
Département génie de la construction, École de technologie
supérieure, Université du Québec, 1100 Notre-Dame Ouest
Montréal (Québec) H3C 1K3, CANADA